Loading…
Ab initio cryo-EM structure modeling that labels density voxels by atom and amino acid type, then threads sequences through them with an HMM.
No providers recorded yet. Browse all providers
A cryo-EM reconstruction is a three-dimensional grid of electron density values, not a structure. Turning it into an atomic model means deciding which voxels hold backbone atoms, which amino acid each belongs to, and in what order they connect along a chain — and with no homologous or predicted structure to lay over the map as a template, all three decisions must come from the density alone.
Cryo2Struct, from Nabin Giri and Jianlin Cheng at the University of Missouri, splits that problem in two. A 3D transformer performs dense voxel classification over the map, labeling every voxel with a backbone atom type and an amino acid type. The predicted alpha-carbons then become the hidden states of a hidden Markov model, with transition probabilities from the spatial distance between candidate atoms and emission probabilities from the network's amino acid predictions; a customized Viterbi search aligns each chain's sequence against that HMM, and the resulting path is the chain-traced backbone. The pipeline is fully automated and calls no external structure predictor.
The transformer never sees sequence — it reads density values only. That is why accuracy barely moves when the test set is filtered to proteins sharing under 25% chain identity with training data, and it is also the limitation the group's later work attacks: both Cryo2Struct2 and MICA push sequence embeddings and predicted structures into the density stream while inheriting this model's alignment machinery.
The architecture follows UNETR: a 32x32x32 single-channel sub-grid is cut into patches of side 16, projected into a 768-dimensional embedding with learned positional encoding, and passed through 12 transformer encoder blocks; features from blocks 3, 6, 9, and 12 feed a convolutional decoder through U-Net skip connections that emits a same-sized labeled grid. The atom-type and amino-acid-type networks are trained separately and hold about 92.3 million parameters each. Training used Cryo2StructData — 6,652 training and 740 validation maps, all deposited before April 2023 — with a batch size of 720, NADAM at 1e-4, dropout 0.1, and a weighted cross-entropy loss, on 24 nodes of six 32 GB NVIDIA V100 GPUs on Summit.
On the 128-map standard benchmark (2.08–5.6 Å, 448–8,416 residues) Cryo2Struct averages 65% alpha-carbon recall against Phenix's 40%, an F1 of 66% against 52%, a length-normalized TM-score of 0.20 against 0.084, and an alpha-carbon quality score of 0.43 against 0.23; its aligned length averages 945.55 residues against 358.51. On a second set of 500 maps deposited after the training cutoff (1.9–4.0 Å, 234–8,828 residues) the figures rise to 70% recall, 70% F1, 0.22 TM-score, 0.50 quality, 49.5% alpha-carbon match, and 20.1% sequence match. Aligned regions average 532.51 residues at 1.6 Å RMSD.
The user is a structural biologist with a reconstructed map and sequences but no reliable template for the chains — common for large multi-chain assemblies and for proteins with no close structural homolog. Output is a chain-traced backbone with residue assignments and per-residue confidence, ready to inspect in ChimeraX and correct by hand. Runtime scales with map size: 2.90 minutes for a 234-residue map, about nine hours for an 8,828-residue one. Phenix real-space refinement of the result raises the alpha-carbon match score by a further 6.5 to 8.9 percentage points.
Published in Nature Communications after peer review, the work answers the question it set out to ask: density alone supports high per-atom recognition — recall and F1 near 70% — but low global TM-scores, because connecting thousands of correctly identified atoms into the right chains and registering the right sequence onto them is combinatorially harder than finding them. That gap set the agenda for the group's follow-ups, which supply the missing sequence and template information rather than more density modeling. The code is MIT-licensed and archived on Zenodo, and the trained checkpoints, benchmark manifests, and modeled structures are on Harvard Dataverse under CC BY 4.0.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.